Detection of small target in infrared images based on multi-band background model

The main Infrared Search and Track systems (IRST) purpose is to realize optimal discrimination between true targets and background clutter (false alarm). In such single band systems, background prediction is frequently used for detecting small targets. However, detection performances are strongly influenced by background gurgitation. The method based on maximum background model can reduce this kind of influence. But present background prediction methods choose background pixels around the prediction pixel from every direction, as a result, background pixels around the target will be 'poisoned' by target, and contrast will be greatly reduced accordingly. Threshold chosen to detect the target in the predicted residual image will decrease, and this will result in too many false targets and increase false alarms. For the small targets detection in IR images, a method of background prediction based on multi-band background model is proposed. For the purpose of removing the target poison, an improved rule of selecting background pixels according to the certain spectral difference between the expected target and background has been developed in this method. The use of this information is based on theoretical spectral radiance discrimination in LWIR and MWIR bands, between targets and backgrounds. When the current spectral parameter matches spectral background response, the current pixel is judged as a background pixel, and involve in background prediction operation, otherwise, it is judged as a target pixel, and will not involve in this operation. The multi-band background model, which improves the performance of small targets detection, eliminates the effect of target on the background prediction, achieves more accurate prediction of background, and increases the contrast of target and background. This is a significant development to the background prediction algorithm by extending to multi-band domain. Simulation results validate the effectiveness of the algorithm in this paper.